*New Paper*
🚨 Goal misgeneralization occurs when AI agents learn the wrong reward function, instead of the human's intended goal.
😇 We show that training with a minimax regret objective provably mitigates it, promoting safer and better-aligned RL policies!
